|
Title:
|
DATA PROCESSING APPROACHES FOR LUNG
CT- IMAGE ANALYSIS UNDER RESOURCE CONSTRAINTS |
|
Author(s):
|
Aleksandra Vatian, Artyom Lobantsev, Nikita Gorokhov, Mikhail Mirzayanov, Georgii Korneev, Natalia Gusarova and Anatoly Shalyto |
|
ISBN:
|
978-989-8533-89-0 |
|
Editors:
|
Mário Macedo |
|
Year:
|
2019 |
|
Edition:
|
Single |
|
Keywords:
|
Lung Tumor, Resource Constraints, Domain Adaptation, Data Preprocessing, CT-Images |
|
Type:
|
Full Paper |
|
First Page:
|
19 |
|
Last Page:
|
26 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
click to dowload 
|
|
Paper Abstract:
|
In this paper we consider the methods of the improvement the efficiency of lung tumor classification made by deep
convolutional networks in conditions of limited resources. We propose two approaches of the efficiency improvenets.
The first is the consideration of the specifics of the brightness distribution of CT images of lung cancer and preprocessing
the data according to these specifics. Experiments with the inclusion of preprocessing proposed by this method showed
an improvement in the classification efficiency on a small amount of computing resources. The second is the use of an
improved network training method for the task of domain adaptation, which includes the integration of a specific loss
function into the network architecture. That loss function aligns the latent feature space of the network trained on source
large dataset of CT scans with the latent feature space of the target small dataset, where the network will be used for
inference. The integration of this method has shown comparable results with the traditional fine-tuning method in a
noticeably smaller amount of time and on the same hardware resources. |
|
|
|
|
|
|
